Method and system for recovering advanced features of low-illumination image based on neural network

Abstract

The invention relates to the technical field of machine vision, and provides a method and a system for recovering advanced features of a low-illumination image based on a neural network. In the method, a low-illumination source image is processed based on a neural network model to obtain a first image, Gaussian blur is performed on the first image and a normal-illumination target image corresponding to the source image through a multi-dimensional Gaussian function, and advanced features of the first image and the target image are extracted; and based on the deformed peak signal-to-noise ratio model, determining differential loss values of advanced features of the first image and the target image, and updating the neural network model so as to perform advanced feature recovery on the low-illumination image based on the updated neural network model. Therefore, in the back propagation process of the neural network model, the neural network model is updated by adding the deformed peak signal-to-noise ratio model into the loss function, so that the neural network model better recovers advanced features in the image.

Description

Technical Field

[0001] The present application relates to the field of image processing technology, and in particular to a method, system, computer-readable storage medium and electronic device for restoring high-level features of low-light images based on a neural network. Background Art

[0002] At present, when restoring the illumination of low-illuminance images, the commonly used methods include image enhancement technology, noise reduction technology and Retinex theory. In image enhancement technology, the commonly used methods include histogram equalization method, contrast-limited adaptive histogram equalization method, gamma correction, etc. These methods can improve the contrast and brightness of low-illuminance images, but will increase noise or distortion. In noise reduction technology, the commonly used methods include wavelet transform, bilateral filter or non-local mean noise reduction, etc., which restore the image by reducing noise. Simple pixel operations will also cause distortion and generate noise.

[0003] The Retinex theory improves image quality by separating the reflection component and the illumination component of the image. However, during the operation, only a single relationship of the illumination between the light rays is considered, and there is only one nonlinear mapping, which will also cause image distortion and generate a large amount of noise.

[0004] Therefore, there is an urgent need to provide a technical solution to the above-mentioned deficiencies in the prior art. Summary of the invention

[0005] The purpose of the present application is to provide a method, system, computer-readable storage medium and electronic device for restoring high-level features of low-light images based on a neural network, so as to solve or alleviate the problems existing in the above-mentioned prior art.

[0006] In order to achieve the above objectives, this application provides the following technical solutions:

[0007] The present application provides a method for restoring high-level features of low-illuminance images based on a neural network, comprising: step S101, processing a captured low-illuminance source image based on a preset neural network model to obtain a first image; step S102, performing Gaussian blur on the first image and a target image of normal illumination obtained corresponding to the source image through a high-dimensional Gaussian function, and extracting high-level features of the first image and high-level features of the target image respectively; step S103, determining differential loss values ​​of the high-level features of the first image and the high-level features of the target image based on a deformed peak signal-to-noise ratio model; step S104, updating the neural network model according to the differential loss value, so as to restore the high-level features of the low-illuminance image based on the updated neural network model.

[0008] Preferably, in step S101, the characteristic illuminations of different levels of entity features in the source image are cyclically extracted to perform convolution operations, and the extracted characteristic illuminations are restored to obtain the first image.

[0009] Preferably, in step S102, based on the multidimensional Gaussian function, Gaussian blur is performed on the first image and the target image respectively to filter out low-level features in the first image and the target image to obtain high-level features of the first image and the target image.

[0010] Preferably, in step S103, according to the formula:

[0011]

[0012] Determine the differential loss value PNSR_DF of the high-level features of the first image and the high-level features of the target image; wherein, is the maximum pixel value of the target image; MSE_DF is the mean square error between the first image and the target image.

[0013] Preferably, in step S103, according to the formula:

[0014]

[0015] Determine the mean square error MSE_DF after deformation;

[0016] Among them, I DF(i,j) is the pixel value of the high-level feature of the first image at coordinate (i, j); K DF(i,j) is the pixel value of the high-level feature of the target image at the coordinate (i, j); n and m are the length and width of the image that only retains the high-level features after high-dimensional Gaussian blurring; i, j, n, and m are all positive numbers.

[0017] Preferably, according to the formula:

[0018]

[0019] Determine the pixel value I of the high-level feature of the first image at coordinates (i, j) DF(i,j) , and the pixel value K of the high-level feature of the target image at coordinate (i, j) DF(i,j) ; Wherein, r is the radius of the Gaussian kernel of the multidimensional Gaussian function; σ is the intensity of Gaussian blur; N1 and N2 are the dimensions of the first image and the target image to be subjected to multidimensional Gaussian blur, respectively, and N1 and N2 are both positive integers.

[0020] Preferably, in step S104, the model parameter gradient of the neural network model corresponding to the differential loss value is calculated; and the model parameters in the neural network model are updated according to the model parameter gradient and the learning rate of the neural network model.

[0021] An embodiment of the present application also provides a system for restoring high-level features of low-illuminance images based on a neural network, comprising: an illumination extraction unit, configured to process a collected low-illuminance source image based on a preset neural network model to obtain a first image; a high-level feature extraction unit, configured to perform a Gaussian model on the first image and a target object with normal illumination acquired corresponding to the source image through a multi-dimensional Gaussian function, and extract high-level features of the first image and high-level features of the target image respectively; a differential calculation unit, configured to determine differential loss values ​​of the high-level features of the first image and the high-level features of the target image based on a deformed peak signal-to-noise ratio model; and a model updating unit, configured to update the neural network model according to the differential loss value, and restore high-level features of the low-illuminance image based on the updated neural network model.

[0022] An embodiment of the present application also provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed, the method for restoring high-level features of low-light images based on a neural network as described above is implemented.

[0023] An embodiment of the present application also provides an electronic device, comprising: a memory, on which is stored a computer program of any of the methods for restoring high-level features of low-light images based on a neural network as described above; and a processor, for retrieving and executing the computer program stored in the memory.

[0024] Beneficial effects:

[0025] The method for restoring high-level features of low-illuminance images based on a neural network provided in an embodiment of the present application is as follows: first, based on a preset neural network model, a collected low-illuminance source image is processed to obtain a first image; and a multi-dimensional Gaussian function is used to perform Gaussian blur on the first image and a target image of normal illumination corresponding to the source image, and high-level features of the first image and high-level features of the target image are extracted respectively; then, based on a deformed peak signal-to-noise ratio model, differential loss values ​​of high-level features of the first image and high-level features of the target image are determined; finally, according to the differential loss values, the neural network model is updated, and high-level features of the low-illuminance image are restored based on the updated neural network model. Thus, in the process of back propagation of the neural network model, the differential loss value of the image is calculated by adding the deformed peak signal-to-noise ratio model to the loss function, and the neural network model is updated, so that the neural network model can better restore high-level features in the image when calculating the target image and the low-illuminance image. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] The drawings in the specification, which constitute a part of the present application, are used to provide further understanding of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute improper limitations on the present application.

[0027] in:

[0028] Figure 1 A flowchart of a method for restoring high-level features of a low-light image based on a neural network according to some embodiments of the present application;

[0029] Figure 2 A schematic diagram of restoring high-level features of a low-light image according to some embodiments of the present application;

[0030] Figure 3 A schematic diagram of restoring high-level features of another low-light image provided according to some embodiments of the present application;

[0031] Figure 4 A schematic diagram of the structure of a system for restoring high-level features of low-light images based on a neural network according to some embodiments of the present application;

[0032] Figure 5 A schematic diagram of the structure of an electronic device provided according to some embodiments of the present application;

[0033] Figure 6 The hardware structure of an electronic device is provided according to some embodiments of the present application. DETAILED DESCRIPTION

[0034] The present application will be described in detail below with reference to the accompanying drawings and in conjunction with embodiments. Each example is provided by way of explanation of the present application and does not limit the present application. In fact, it will be clear to those skilled in the art that modifications and variations may be made in the present application without departing from the scope or spirit of the present application. For example, a feature shown or described as a part of an embodiment may be used in another embodiment to produce yet another embodiment. Therefore, it is desired that the present application includes such modifications and variations within the scope of the appended claims and their equivalents.

[0035] In the field of machine learning and deep learning, processing low-light images is an important research direction, which aims to improve the visibility, details and color performance of images so that they can be displayed clearly under low-light conditions. Deep neural networks are used to automatically learn the mapping relationship from low-light images to high-light images, automatically repair images, and adjust the brightness and color of images. However, whether it is image enhancement technology, noise reduction technology or Retinex theory, when restoring the illumination of the image, it will cause distortion and generate noise.

[0036] Based on this, an embodiment of the present application provides a method for restoring high-level features of low-light images based on a neural network. During the back propagation process of training a neural network model (convolutional neural network or adversarial neural network), the deformed peak signal-to-noise ratio is added to the calculated loss function, thereby guiding the neural network model to better restore the high-level features of the image.

[0037] like Figures 1 to 3 As shown, the method for restoring high-level features of low-light images based on a neural network includes:

[0038] Step S101: Based on a preset neural network model, a collected low-illuminance source image is processed to obtain a first image.

[0039] Low-light images refer to images with insufficient illumination (light intensity), which appear dark to the human eye. In this application, images with an average brightness below 50 lux, a peak brightness below 80 lux, an image contrast below 30:1, and a color saturation below 70% are defined as low-light images.

[0040] The source image and the target image are paired images of the same scene, captured under low illumination conditions and normal illumination respectively. The training of the neural network model for restoring image illumination uses paired source images and target images to ensure that the training network can learn the mapping from low illumination conditions to normal illumination conditions.

[0041] In the present application, when processing the collected low-illuminance source image, the first image is obtained by cyclically extracting the characteristic illumination of different levels of the entity features in the source image, performing a convolution operation, and restoring the extracted characteristic illumination. That is to say, in the neural network model, the characteristic illumination of different levels of the entity features in the source image is continuously extracted through the convolution operation, and then the extracted characteristic illumination of different levels is again convolved to restore the extracted characteristic illumination of different levels. Since the illumination of the high-level features of the entity features is restored in the loss function of the neural network model, the second convolution operation only restores the illumination of the high-level features under the guidance of the loss function. Here, the high-level features of the image include but are not limited to the contour, material, shape, size, relative position relationship between the entity features (objects), and characteristic functions of the entity features (objects) in the image.

[0042] Step S102: Gaussian blur the first image and the target image with normal illumination obtained and corresponding to the source image by using a multi-dimensional Gaussian function, and extract high-level features of the first image and high-level features of the target image respectively.

[0043] In the present application, the target image and the source image are paired images of different illuminations captured in the same scene, wherein the target image is an image captured under normal illumination conditions, and the source image is an image captured under low illumination conditions. When obtaining the high-level features of the first image and the high-level features of the target image, based on a multidimensional Gaussian function, Gaussian blur is performed on the first image and the target image, respectively, and low-level features in the first image and the target image are filtered out to obtain high-level features of the first image and the target image. That is, the first image and the target image are filtered by a multidimensional Gaussian function, respectively, and the high-frequency low-level features in the image are filtered out, and the high-level features in the image are retained, thereby realizing the extraction of high-level features of the image.

[0044] That is, first, a multidimensional Gaussian function is obtained by distributing weights of the distances between pixels in the image and the central pixel; then, the multidimensional Gaussian function is used as a weight coefficient to reduce high-frequency components (i.e., details, noise) in the first image and the target image, and filter out low-level features in the first image and the target image. Here, the low-level features of the image include but are not limited to pixels, saturation, brightness, hue, and lightness in the image.

[0045] In the Gaussian blur process, a Gaussian kernel (two-dimensional array) is created based on a multidimensional Gaussian function. Then, for each pixel in each color channel of the image, the Gaussian kernel is used to perform weighted averaging of the pixels in its neighborhood. The value at each pixel position is replaced with the weighted average, thereby blurring the image.

[0046] Step S103: Determine the differential loss values ​​of the high-level features of the first image and the high-level features of the target image based on the deformed peak signal-to-noise ratio model.

[0047] In this application, the difference between the high-level features of the first image and the high-level features of the target image is calculated by the peak signal-to-noise ratio, which is the differential loss value between the first image and the target image. Specifically, according to the formula:

[0048]

[0049] Determine a differential loss value of the high-level features of the first image and the high-level features of the target image.

[0050] in, is the maximum pixel value of the target image; MSE_DF is the mean square error after deformation, which is used to determine the L2 norm from the first image to the target image.

[0051] Here, according to the formula:

[0052]

[0053] Determine the mean square error MSE_DF between the first image and the target image.

[0054]

[0055] In the formula, I DF(i,j) is the pixel value of the high-level feature of the first image at coordinate (j, j); K DF(i,j) is the pixel value of the high-level feature of the target image at coordinates (i, j); n and m are the length and width of the image (first image, target image) that only retains the high-level features after multi-dimensional Gaussian blurring; i, j, n, and m are all positive numbers; r is the radius of the Gaussian kernel of the multi-dimensional Gaussian function; σ is the intensity of the Gaussian blurring; N1 and N2 are the dimensions of the first image and target image to be subjected to multi-dimensional Gaussian blurring, respectively, and N1 and N2 are both positive integers.

[0056] Step S104: updating the neural network model according to the differential loss value, and restoring the high-level features of the low-illumination image based on the updated neural network model.

[0057] In the present application, the back propagation of the differential loss value, specifically the back propagation of the convolutional neural network or the adversarial neural network, is used to guide the neural network (convolutional neural network or adversarial neural network) model to learn to restore the high-level features of the low-light image. Specifically, first, the model parameter gradient of the neural network model corresponding to the differential loss value is calculated; then, according to the model parameter gradient and the learning rate of the neural network model, the model parameters in the neural network model are updated, thereby guiding the neural network model to learn to restore the high-level features of the low-light image.

[0058] Here, the differential loss value between the high-level features of the first image and the high-level features of the target image is calculated by using the deformed peak signal-to-noise ratio model, and the differential loss value is used as the loss function, that is, the deformed peak signal-to-noise ratio is added to the original loss function of the neural network model. Then, the gradient of the deformed loss function (differential loss value) relative to the model parameters of the neural network model is calculated; finally, the direction and size of the gradient descent of the model parameters are obtained by multiplying the gradient by the learning rate of the neural network model, so as to update the model parameters. That is, the guidance weight of the loss function in the back propagation process is updated based on the gradient descent method.

[0059] The neural network model accepts input data and passes it through a hierarchical structure, with each layer calculating its output until the last layer generates the final prediction result; the loss function calculates the difference between the predicted result and the actual label, generating a single loss value (differential loss value) to measure the performance of the current network.

[0060] The gradient of the loss function with respect to the model parameters (guidance weights) of the neural network model is calculated through the back propagation network. Specifically, it is achieved by differentiating the network layer by layer, using the chain rule from the output layer to the input layer. Among them, the gradient represents the rate of change of the differential loss value relative to each weight.

[0061] Then, using the gradient descent method or other optimization algorithms, the guidance weights are updated according to the calculated gradients to reduce the value of the loss function. Usually, the update of the guidance weights is usually the guidance weight minus the product of the learning rate and the gradient, and the difference is the updated guidance weight. Among them, the learning rate is a hyperparameter that determines the magnitude of each update in the optimization process.

[0062] Through iterative cycles and continuous optimization of the guidance weights, the prediction results of the neural network model will gradually approach the true label, and the value of the loss function will decrease accordingly. In this process, the loss function provides a standard for measuring the performance of the model, and guides the update direction and magnitude of the weights through its gradient to train the model to make accurate predictions.

[0063] The neural network model with updated model parameters can be used to restore high-level features of low-light images. Here, the high-level features of the image include but are not limited to the contours of objects in the image, the materials of objects, the colors of objects, the shapes of objects, the sizes of objects, the relative positions between objects, and the characteristic functions of objects.

[0064] In the present application, the acquired low-illuminance source image is subjected to illumination restoration through a convolutional neural network or an adversarial neural network to obtain a first image; then, the first image and the target image are subjected to multi-dimensional Gaussian blur to extract high-level features of the first image and the target image, and the difference (differential loss value) between the high-level features of the first image and the target image is calculated through a deformed peak signal-to-noise ratio model. Through back propagation, the convolutional neural network or the adversarial neural network is guided to focus on restoring the high-level features of the low-illuminance image, rather than simply restoring the illumination between pixels and pixel groups highlighted by low illumination, so that the illumination restoration effect of the low-illuminance image is better.

[0065] like Figure 4 As shown, the embodiment of the present application also provides a system for restoring high-level features of low-light images based on a neural network, including:

[0066] The illumination extraction unit 401 is configured to process the collected low-illuminance source image based on a preset neural network model to obtain a first image;

[0067] The high-level feature extraction unit 402 is configured to perform Gaussian modeling on the first image and the target object with normal illumination obtained corresponding to the source image through a multi-dimensional Gaussian function, and extract high-level features of the first image and high-level features of the target image respectively;

[0068] A differential calculation unit 403 is configured to determine differential loss values ​​of high-level features of the first image and high-level features of the target image based on the deformed peak signal-to-noise ratio model;

[0069] The model updating unit 404 is configured to update the neural network model according to the differential loss value, and perform high-level feature recovery on the low-illuminance image based on the updated neural network model.

[0070] The system for restoring high-level features of low-light images based on a neural network provided in the embodiments of the present application can implement the steps and processes of any of the above-mentioned method embodiments for restoring high-level features of low-light images based on a neural network, and achieve the same technical effects, which will not be described one by one here.

[0071] Figure 5 Schematic diagram of the structure of an electronic device provided according to some embodiments of the present application; Figure 5 As shown, the electronic device includes:

[0072] One or more processors 501;

[0073] The computer-readable medium can be configured to store one or more programs 502. When one or more processors 501 execute one or more programs 502, the following steps are implemented: based on a preset neural network model, the collected low-illuminance source image is processed to obtain a first image; through a multi-dimensional Gaussian function, the first image and the acquired target image with normal illumination corresponding to the source image are Gaussian blurred to extract high-level features of the first image and high-level features of the target image respectively; based on the deformed peak signal-to-noise ratio model, the differential loss values ​​of the high-level features of the first image and the high-level features of the target image are determined; according to the differential loss values, the neural network model is updated, and the high-level features of the low-illuminance image are restored based on the updated neural network model.

[0074] Figure 6 The hardware structure of the electronic device provided according to some embodiments of the present application; Figure 6 As shown, the hardware structure of the electronic device may include: a processor 601 , a communication interface 602 , a computer-readable medium 603 and a communication bus 604 .

[0075] The processor 601 , the communication interface 602 , and the computer-readable storage medium 603 communicate with each other via a communication bus 604 .

[0076] Optionally, the communication interface 602 may be an interface of a communication module, such as an interface of a GSM module.

[0077] Among them, the processor 601 can be specifically configured as: based on a preset neural network model, processing the collected low-illuminance source image to obtain a first image; through a multi-dimensional Gaussian function, Gaussian blurring the first image and the acquired target image of normal illumination corresponding to the source image, respectively extracting the high-level features of the first image and the high-level features of the target image; based on the deformed peak signal-to-noise ratio model, determining the differential loss values ​​of the high-level features of the first image and the high-level features of the target image; according to the differential loss values, updating the neural network model, and restoring the high-level features of the low-illuminance image based on the updated neural network model.

[0078] The processor 601 may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc. It may also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components. It may implement or execute the disclosed methods, steps and logic block diagrams in the embodiments of the present application. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0079] The electronic device of the embodiment of the present application exists in various forms, including but not limited to:

[0080] (1) Mobile communication devices: These devices are characterized by their mobile communication functions and their main purpose is to provide voice and data communications. These terminals include: smart phones (e.g., iPhone), multimedia phones, functional phones, and low-end phones.

[0081] (2) Ultra-mobile personal computer devices: These devices fall into the category of personal computers, have computing and processing capabilities, and generally also have mobile Internet access features. These terminals include: PDA, MID and UMPC devices, such as iPad.

[0082] (3) Portable entertainment devices: These devices can display and play multimedia content. They include audio and video players (such as iPods), handheld game consoles, e-books, smart toys, and portable car navigation devices.

[0083] (4) Server: A device that provides computing services. The server consists of a processor, hard disk, memory, system bus, etc. The server is similar to a general computer architecture, but because it needs to provide highly reliable services, it has higher requirements in terms of processing power, stability, reliability, security, scalability, and manageability.

[0084] (5) Other electronic devices with data interaction functions.

[0085] It should be pointed out that, according to the needs of implementation, the various components / steps described in the embodiments of the present application can be split into more components / steps, or two or more components / steps or partial operations of components / steps can be combined into new components / steps to achieve the purpose of the embodiments of the present application.

[0086] The above-mentioned method according to the embodiment of the present application can be implemented in hardware, firmware, or implemented as software or computer code that can be stored in a recording medium (such as a CD ROM, RAM, floppy disk, hard disk or magneto-optical disk), or implemented as computer code originally stored in a remote recording medium or a non-temporary machine storage medium downloaded through a network and to be stored in a local recording medium, so that the method described herein can be stored in such software processing on a recording medium using a general-purpose computer, a dedicated processor or programmable or dedicated hardware (such as an ASIC or FPGA). It can be understood that a computer, a processor, a microprocessor controller or programmable hardware includes a storage component (e.g., RAM, ROM, flash memory, etc.) that can store or receive software or computer code. When the software or computer code is accessed and executed by a computer, a processor or hardware, the method described herein for restoring high-level features of low-light images based on a neural network is implemented. In addition, when a general-purpose computer accesses the code for implementing the method shown herein, the execution of the code converts the general-purpose computer into a dedicated computer for executing the method shown herein.

[0087] Those of ordinary skill in the art will appreciate that the units and method steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and constraints involved in the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the embodiments of the present application.

[0088] It should be noted that the various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device and system embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiments.

[0089] The device and system embodiments described above are merely illustrative, wherein the units not shown as separate may or may not be physically separated, and the units not shown as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Those of ordinary skill in the art may understand and implement it without creative work.

[0090] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

[0091] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for restoring high-level features of low-light images based on neural networks, characterized in that: include: Step S101: Based on a preset neural network model, the collected low-illumination source image is processed to obtain a first image; Step S102: Gaussian blurring the first image and the acquired target image with normal illumination corresponding to the source image by using a high-dimensional Gaussian function, and extracting high-level features of the first image and high-level features of the target image respectively; Step S103: determining differential loss values ​​of high-level features of the first image and high-level features of the target image based on the deformed peak signal-to-noise ratio model; Step S104: updating the neural network model according to the differential loss value, so as to restore high-level features of the low-illuminance image based on the updated neural network model.

2. The method for restoring high-level features of low-light images based on a neural network according to claim 1, characterized in that: In step S101, The characteristic illuminations of different levels of entity features in the source image are cyclically extracted to perform convolution operations, and the extracted characteristic illuminations are restored to obtain the first image.

3. The method for restoring high-level features of low-light images based on a neural network according to claim 1, characterized in that: In step S102, Based on the multidimensional Gaussian function, Gaussian blur is performed on the first image and the target image respectively to filter out low-level features in the first image and the target image, so as to obtain high-level features of the first image and the target image.

4. The method for restoring high-level features of low-light images based on a neural network according to claim 1, characterized in that: In step S103, according to the formula: Determine a differential loss value PNSR_DF of the high-level features of the first image and the high-level features of the target image; in, is the maximum pixel value of the target image; MSE_DF is the mean square error between the first image and the target image.

5. The method for restoring high-level features of low-light images based on a neural network according to claim 4, characterized in that: In step S103, according to the formula: Determine the mean square error MSE_DF after deformation; Among them, I DF(i,j) is the pixel value of the high-level feature of the first image at coordinates (i, j); K DF(i,j) is the pixel value of the high-level feature of the target image at coordinate (i, j); n and m are the length and width of the image after high-dimensional Gaussian blurring, which only retains high-level features; i, j, n, and m are all positive numbers.

6. The method for restoring high-level features of low-light images based on a neural network according to claim 5, characterized in that: According to the formula: Determine the pixel value I of the high-level feature of the first image at coordinates (i, j) DF(i,j) , and the pixel value K of the high-level feature of the target image at coordinate (i, j) DF(i,j) ; Among them, r is the radius of the Gaussian kernel of the multidimensional Gaussian function; σ is the intensity of Gaussian blur; N1 and N2 are the dimensions of the first image and the target image to be subjected to multidimensional Gaussian blur, respectively, and N1 and N2 are both positive integers.

7. The method for restoring high-level features of low-light images based on a neural network according to claim 1, characterized in that: In step S104, Calculating the model parameter gradient of the neural network model corresponding to the differential loss value; The model parameters in the neural network model are updated according to the model parameter gradient and the learning rate of the neural network model.

8. A system for restoring high-level features of low-light images based on neural networks, characterized in that: include: an illumination extraction unit, configured to process the collected low-illuminance source image based on a preset neural network model to obtain a first image; a high-level feature extraction unit configured to perform a Gaussian model on the first image and the target object with normal illumination obtained corresponding to the source image through a multi-dimensional Gaussian function, and extract high-level features of the first image and high-level features of the target image respectively; A differential calculation unit configured to determine differential loss values ​​of high-level features of the first image and high-level features of the target image based on the deformed peak signal-to-noise ratio model; A model updating unit is configured to update the neural network model according to the differential loss value, and perform high-level feature restoration on the low-illuminance image based on the updated neural network model.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed, the method for restoring high-level features of low-light images based on a neural network as described in any one of claims 1 to 7 is implemented.

10. An electronic device, characterized in that: include: A memory storing a computer program of the method for restoring high-level features of a low-light image based on a neural network as claimed in any one of claims 1 to 7; A processor is used to retrieve and execute the computer program stored in the memory.

Images